After introduced and explained by Geoffrey Hinton’s paper published in 2006, deep learning algorithm has been applied in improving the accuracy of image recognition by computer scientists these years. In the past of decade, scientists are continually making breakthrough on the image-related technologies, including image recognition, segmentation, character extraction, quantitative analysis and comparative analysis. These advanced technologies has been applied in various industries, especially in the works relying on images originally. Medical imaging diagnosis is a typical one.

Medical image, including many types like X-ray, computed tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI), etc., is used to diagnose many diseases by doctors today. However, the number of Chinese imaging doctors is in short supply, and there is a shortage of doctors with abundant clinical experience, especially in primary hospitals in towns and villages. In addition, doctors diagnose mainly relying on their experience, so there are lots of misdiagnosis and missed diagnosis.

Since 2012, many start-ups have emerged in China, which apply image recognition and other image-related techs in the process of image diagnosis. They hire top-notch artificial intelligence engineers and outstanding imaging doctors to train algorithmic models and develop medical image-assisted diagnostic products. These products can greatly improve diagnostic efficiency and diagnostic accuracy, which are really good assistants for doctors.

However, these products are forbidden to commercialize without the certification from CFDA (China Food and Drug Administration). In other words, if require the certification, the products can enter the hospital’s medical suppliers list and can be purchased and reimbursed by national medical insurance fund. Because these AI Medical Imaging Diagnosis Products are newly emerging ones, CFDA does not have mature certification standards, so the assessment progress is really slow. For now, there are only two China’s companies requiring the certifications. So how do these companies run their business and where does income come from?

Medical valuation of AI medical imaging diagnosis

AI medical imaging diagnosis products emerged in China are mainly for cancers，especially lung cancer and breast cancer. There are a large number of cancer patients in China. More than 10,000 people were diagnosed with cancer every day in 2014, which means that 7 people were diagnosed with cancer every minute, of which lung cancer ranked first, according to the National Cancer Center. In 2017, the number of lung cancer patients reached 800,000, and the deaths were nearly 700,000. And the death units are still growing at a rate of 26.9% per year. It is estimated that there will be 1 million people dying from lung cancer in China in 2025, according to the Capital Medical University.

Breast cancer is the most severe malignant tumor among female population. According to the cancer report published by the World Health Organization's International Agency, which contains the research results from 184 countries and regions, the incidence and mortality rate of breast cancer among Chinese women is at a low level in the world, but the morbidity and mortality have increased rapidly in the past 20 years. In addition, the number of Chinese breast cancer patients accounts for 11.19% of the world. According to the National Cancer Center, there were 278,900 new breast cancer patients emerged in 2014, accounting for 16.51% of the incidence of female malignant tumors, which ranks first in the incidence of female malignant tumors.

So how many imaging doctors does China have? The answer is only 150 thousand. Besides, more than 50% of imaging doctors work more than 8 hours a day, and 20.6% of them work more than 10 hours a day. The annual growth rate of medical imaging data in China is about 30%, while the annual growth rate of imaging doctors is only 4.1%, according to VCBeat.net, a medical content platform. Under the heavy workload, doctors are probably to make mistake on diagnosis.

DeepCare, a Chinese start-up developing related products, made an experiment to analyze the performance between doctors with different working length. In this experiment, a pathologist with 40-year working length and a pathologist with 10-year working length diagnosed the same group of lymphatic metastasis digital pathological sections. The result shows that the disparity of true positive rate (TPR) between these two doctors are as high as 30%.

Using AI to help doctors make diagnosis can greatly reduce the workload of doctors and improve the accuracy. For example, InferVsion (推想科技), a China’s start-up developing AI-enabled medical imaging diagnosis solutions, provided its two products, AI-DR and AI-CT, to Tongji Medical College of HUST. These products detected lung cancer from 110,000 X-rays and more than 3,000 CTs, and the detection rate surpassed 92% and 95% respectively with only 5 seconds.

Landscape of AI medical imaging diagnosis market

So how many companies develops these kind of AI products? According to EO Company, there are 61 typical companies in China, and 41 are founded between 2014 and 2017. Since 2016, investments in China's AI medical enterprise has begun to explode. The number of investments has increased from 10 in 2015 to 22 in 2016, while the investment amount has increased from CNY 162 million (USD 23.75 million) in 2015 to CNY 1291 million (USD 189.25 million) in 2016. The peak of investments occurred in 2017, when there are 35 investment totally in CNY 4.95 billion (USD 0.73 billion).

It is said by some investors that there are more than 100 companies developing this kind of products in China, but some companies are not known to the public. Among these 100 companies, 90% of them develop diagnostic products for lung cancer, because the number of cancer patients in China is huge, which means the lung cancer case data is rich. And the algorithm models need a large amount of data for training. Besides, there are abundant papers and academic materials to analyze lung cancer, and the lung cancer’s research is the most mature among various types of cancer.

In addition to lung cancer, there are also some products for other diseases, such as diabetes, stroke, breast cancer, esophageal cancer, fractures, skin diseases, liver cancer, prostate cancer, and cardiovascular diseases. Representative companies in this field include inferVision, DeepWise (深睿医疗), 12Sigma (图玛深维), YITU Medical (依图医疗), VoxelCloud (体素科技), United Imaging Healthcare (联影医疗), Airdoc, etc.

InferVision was founded in 2015, and it is the top enterprise in this field. It has raised for 5 rounds of financing totally in more than CNY 500 million (USD 73.30 miilion). Its core product namely InferRead CT Lung, which is extremely sensitive to tiny nodules and ground glass nodules. It can help doctors improve diagnostic efficiency by 30%-50% with ensuring accuracy. As of May in 2018, the product has been adopted in more than 150 China’s hospitals, mainly in the first-class hospitals.

DeepWise was founded in 2017, a start-up with rapidly development. It has raised 3 rounds of financing within one year since it founded, totally in nearly CNY 300 million (USD 43.98 million). The company's core product is also for cancer diagnosis, and it only takes 30 seconds to make a diagnosis. The accuracy of this product reaches 98.8%, which means its sensitivity and specificity of lung nodule detection have reached the international leading level. As of December 2018, this product has been adopted in nearly 300 China’s hospitals.

12Sigma was founded in 2016, and it has raised 4 rounds of financing totally in about CNY 250 million (USD 26.65 million). Its core product namely is σ-Discover-Lung, which is also for the lung nodule detection and analysis. In addition to leveraged by deep learning algorithm, this product is also improved by the 3D segmentation technology independently developed by 12Sigma. As of June 2018, this product has been adopted to more than 100 China’s first-class hospitals.

YITU Medical was founded in 2012, which is a subsidiary of YITU, a China's AI unicorn enterprise. Its core product, Care.Ai TM, is for a variety of diseases, including lung cancer, bone age detection, pediatric intelligent diagnosis, and medical record intelligent search engine. As of January 2019, more than 50 top hospitals in China has adopted this product.

However, there are only two China’s companies obtaining the certification from CFDA. Without this certification, products cannot enter the hospital's procurement list. On August 1 of 2018, the new edition of the Catalogue of Medical Devices, which was updated by CFDA, was officially put into effect. Among them, the category of "in vitro diagnostic software" contains the AI medical imaging diagnosis products. According to this category, if the diagnostic software provides diagnostic recommendations through its algorithm with only has the auxiliary diagnostic function - does not directly give the diagnosis conclusion - the related products are managed as the Class II of medical devices. If the software automatically identifies the lesion by its algorithm and provides clear diagnostic prompts, the software’s risk is relatively high, and the related products are managed as the Class III of medical device.

The third type of medical device application is more difficult than the second type, in which clinical trials are needed for the third type. Till now, only Landing-Med (兰丁医学) and EDDA Technology (医软信息科技) have obtained CFDA certification, so the two companies have already bid farewell to the free trial stage, and customers can pay them legally. So how do other companies that have not been certified run business? In fact, companies and investors are trying to explore a new business model. At present, these companies earn money by cooperating with doctors in scientific research. Enterprises provide doctors with support for algorithm models, computing sources, data processing and analysis. And companies can obtain a part of the compensation from research funding. Of course, another more important value of scientific research cooperation is to cooperate with hospitals and use the hospital case data to train their algorithm models.

Most companies mentioned above are applying for a Class II certificate, but there is no new progress. According to industry investors, although the new edition of the Catalogue of Medical Devices has been in operation for more than half a year, the certification system for AI imaging diagnosis products is still being improved. It is expected that this kind of products will successfully apply for Class II certificates by 2020, and certificated products should be hardware combining with software.